Machine Learning and Friends Lunch
Toward better face recognition: non-parametric maximum likelihood methods for separating causes in images
Erik Learned-Miller
UMass
Abstract
Consider the problem of face recognition. Why is it so hard? One reason is that there are so many causes of variability in the image of a face: pose, lighting, camera parameters, facial expressions, and bad-hair days to name a few. If we can separate the causes of appearance into those related to the identity of a person and those that are simply "noise", then face recognition should be easier. But this is quite a tall order.
In this talk, I will discuss methods for separating types of variability in image data sets. I will start with the idea of removing shape variability from handwritten digits and show how this leads to new types of models for recognizing handwritten digits. Then I discuss removing certain types of variability from MR medical images, allowing a more accurate analysis of the underlying tissues. I will try to convince you that if we can continue to separate types of variability in images, eventually we will be able to produce models for faces that will enable human quality face recognition.